What is artificial intelligence, asks Niklas Kuhl, head of the Applied AI Lab, Karlsruhe Institute of Technology/IBM. He continues: “To answer that question, you need to ask, what is intelligence? One possible definition is the ability to perceive information, to retain it as knowledge and to apply it towards adaptive behaviours within an environment. This applies to synthetic and human intelligence. There are different tests for intelligence. You can plot a graph and arrange living beings from goldfish to mouse, dog, whale, monkey, human. This is the biological range of intelligence which we can observe on the planet.
“The first definition of AI is weak AI. It is good at one particular thing: playing Go, playing chess, predicting whether a part in your production will be faulty, but it is no good at anything else. The next level of AI is Strong AI. That is an artificial agent which is as intelligent as a human in all areas: this includes social skills, creativity, wisdom, knowledge. In the scale of human intelligence, just as much has a human. Then we have a range of what we call artificial superintelligence, ranging from just a bit smarter than the average human on the planet up to that of many planets combined. Artificial superintelligence is much smarter than the average human brain at every field, including scientific creativity, wisdom and social skills.”
NOT NEW
AI is not a new term, points out Dr. Johannes Kunze von Bischhoffshausen, director of digital transformation at Trelleborg. He says: “With the rise of computers, there was early excitement about whether computers could mimic humans. Even in the 1950s, one of the computer science pioneers Alan Turing came up with the idea that there must be a test to see if AI is smart enough when a human cannot distinguish between human and computer: a kind of chat bot. But then computing power was very limited, so it was a lot of work to program these chatbots.
“So in the 1970s-80s they went in a different direction: the research area of machine learning. In general, that means computers learn from the data itself instead of being explicitly programmed. That was a powerful approach, because then the effort to train systems – at the time people were talking about expert systems – was much less because computers were able to learn themselves. But in the 80s and 90s data wasn’t there in the amounts that we have today with the rise of the internet, and of course computing power wasn’t there either. Only with the rise of distributed computing, or as we call it today, the internet of things, and much more computing power we see a specific area booming which is called deep learning – autonomous driving can rely on them and only then do we see artificial intelligence exploding.
Continues Kuhl: “The change of mindset is that we don’t explicitly program an agent to do something, but that it learns from patterns that it observes in data. AI relies on machine learning. There are three important types of machine learning. The first type is supervised. As a father, I will take some real-life examples from my family. If you give your child a set of toy cars and explain that there are different types of cars: SUVs, sports cars, vans, and so on, you give him or her a bunch of training data. What the child does is learn what the patterns are in the data with the labels that you gave – the distinct characteristics of sports cars and suvs. If you give your child a new bunch of unlabelled cars, then your child is able to map them into these classes that you gave it already. Typical applications in learning terms is regression or classification.
“A different approach is unsupervised. You give your children a set of toy cars and tell them to sort them, but don’t tell them anything. Then they will find their own patterns, for example, my child will sort these cars by colour or amount of wheels or by shape. Unsupervised learning is about finding previously unidentified patterns in the data, and typical learning examples are clustering or association rules.
“Up until 10 years ago, these would have been the only ones. A newer one is reinforcement learning. The difference here is that you get feedback or rewards based on actions taken. If a child picks a toy and labels it, then it will get a reward or a punishment. By doing that, he or she learns over time. You always have an interaction between an action and a reward or punishment. That’s an important driver.”
APPLICATIONS OF AI
von Bischhoffshausen goes on focus on the uses of this technology. He says: “One of the big applications of Internet of Things in general is smart home. AI and IOT are fitting together nicely here. The idea of a smart home is by interconnecting as many sensors and capturing as much data as possible, learning about the consumption patterns, and then afterwards for example optimising the time when the washing machine runs. This is quite capital-intensive. A lot of investment is required to put sensors everywhere.
“But there is a different approach where you just use one signal on the power line, and indirectly learn by means of AI what is the specific pattern of a device by switching on and off, and therefore indirectly measure energy consumption. That is just one example where you can see that actually AI can replace sensors and can reduce amount of physical sensors required in smart home.
“The question for us is, what are the real use cases of AI in industry? Before we show some examples, there is one thing that is important to mention. The value of the use cases are very industry-specific. There is a very nice study from McKinsey [https://is.gd/cotecu] about the different branches and use cases of AI that highlights this. It compares the potential impact of AI on the retail industry to that of the industrial world. The bubble indicates the potential impact, and in retail it is run by marketing and sales. But in industrial and automotive, the value is in supply chain and manufacturing, and according to this study, the biggest area of potential benefit is predictive maintenance.”
Adds Kuhl: “Another example of AI from industry, is from companies’ incident ticket systems. Whenever a customer has trouble with software or hardware, they can submit a ticket. You probably know these. We thought that there was huge potential to not only use tickets to solve the customer’s particular problem, but to identify whether this particular customer is satisfied with the company. There is a flavour to the tickets that they write. We had a lot of data from a ticketing system, and a lot of data from a customer satisfaction survey. So with supervised machine learning, we could learn in what cases customers were satisfied, and in which cases they weren’t satisfied. This is extremely helpful, because it means businesses can execute countermeasures in their own processes early on to increase customer satisfaction so customers don’t churn. You can utilise this point of contact that you have, once the customer engages with you, to make sure he or she is still happy with the service.”
von Bischhoffshausen lists five broad areas where Trelleborg applies AI: in smart manufacturing; improving efficiency in manufacturing; supporting engineering activities; creating new services; product/application. He offers two specific examples in manufacturing. “First is cross-process analysis. The big value is created when you are not just looking at one dataset, for example the data of an injection moulding machine, but actually close the loop and connect all relevant information – raw material, primary production, injection moulding, secondary operations and finally quality data. There is also a big value driver called predictive quality that helps us improve our efficiency, to reduce scrap and contribute to sustainability. The second area of AI in manufacturing is supporting local automation: automating things that would not have been possible before. One potential application is image recognition; detecting quality defects instead of a manual inspection. That is possible with AI on parts where it has not been possible before.”
Kuhl says that KIT worked with Trelleborg to monitor some of its products in use, such as an aeroplane gearbox hydraulics seal. He explains: “But if you do put a sensor into the seal, this changes the behaviour of the seal, and is also not always economical to do so. What we found is that we could use the environment around this seal, where there are already existing sensors in place, to triangulate the status of the seal, with machine learning.” In other words, the combination of temperature, vibration, pressure and other indicators is interpreted by the system to obtain a picture of the state of the sealing system.
Adds von Bischhoffshausen: “Over the years we took this from an academic exercise on our rigs to real customer applications to support their predictive maintenance activities.” He explains that this approach to maintenance is in contrast to the two standard ways - corrective, that means you fix it when it breaks - and preventative, that means that maintenance happens too early. “With predictive maintenance, you change a part when it is exactly required. You need to have the right sensor signals. It can be ultrasonic, vibration, temperature, pressure. But the question is really, what do these signals mean to your system. Our approach to interpret these sensor signals is called ‘cognitive sealing’. We have now more than 10 projects in this area where we apply AI.”
This article is based on the webinar Artificial Intelligence - Applications in Daily Life, Industry & Sealing Technology” (https://is.gd/agolip) first presented in October 2020.
BOX: Is AI only relevant for big businesses?
Niklas Kuhl: “Absolutely they should invest because it could become their core competitive advantage. Typically small to medium-sized businesses are much more flexible, dynamic and faster to integrate innovative ideas into their processes. As of now, the implementation of AI, if done correctly, can give a huge benefit, compared to competitors. AI is definitely not only for large companies but is especially interesting to SMEs.”
Johannes Kunze von Bischhoffshausen: “What is important is what an SME does with it; they do not have infinite resources. Focusing on the right use cases and ‘what do I want to do internally?’ and ‘what do I want to outsource?’ The third option, which has not often been done before, is partnership with suppliers, or research instutitutes. Having a strategic plan for how to integrate AI into the business is crucial – and that applies as much to SMEs as large organisations.”
BOX: Will AI overtake engineering knowledge?
Niklas Kuhl: “I would say not. What we see in research is that there are two ideas of AI. One is that AI will never be able to replace human beings and the tasks that they do; the human will always be the enlightened one who has the most process knowledge and makes the decisions. The other idea is that AI will completely replace humans and outsmart them pretty soon. But in the current research, nearly everyone says that you need to have something called hybrid intelligence: a collaboration between humans and machines where both can play out their strengths. You can automate tedious tasks that typically humans do not like to do to AI, and the more creative tasks can be done by humans. If you think about the knowledge of global engineering, it will be shared between humans and AI, and they will do their best in collaboratively work together.”
Johannes Kunze von Bischhoffshausen: “We see in some applications also in Trelleborg that AI can help to automate simple tasks and replace all of the trial-and-error that engineers are doing. There we see even in the near future that AI helps engineers to focus on the right things. On the other hand, what does this mean for organisations? Hybrid intelligence only works if engineers understand what the AI does and what the results are. Specifically, when you talk about machine learning and its results, there is never 100% correct answer. People in general dealing with AI need to understand how those systems work in order to judge what those recommendations are and what to do with them. It’s quite important that companies not just have a small team of experts but teach their employees how to use AI and how to interpret the results.
BOX: How far has AI developed in Europe, compared to the USA and Asia?
Niklas Kuhl: “There are two sides to this question. One is, what is publicly available – what do you know about the research being done. How much of this is published and available? If you look for example into companies like Facebook and Apple, they don’t reveal all of their techniques and their findings, so it’s hard to say where we are on that scale. If you look into the amount of published research papers, then the US is dominating the market, but if you really want to dig deeper, it’s hard, because you don’t know what exists in these tech companies. If I would have to answer, Id say that the US is deintely leading while Asia is picking up momentum while we in Europe have to intensify our investments into this area.
Johannes Kunze von Bischhoffshausen: “It heavily depends on the use case. When we look into industrial automation or autonomous driving you will get a different picture than just web analytics, for example. It depends on the application.”